Fine-grained Sentiment Analysis Based On Deep Learning

With the rapid development of the Internet,more and more ordinary netizens can express their preferences and attitudes to things through online platforms,such as product reviews,movie reviews,and so on.These reviews contain a lot of valuable information.On the one hand,consumers can understand the reputation of the product through product reviews,and then make corresponding purchase decisions;on the other hand,manufacturers can use the reviews to discover the problems of the products and improve the products.quality.Emotional analysis is an important research direction of natural language processing.The application of emotion analysis technology to the processing and analysis of a large amount of unstructured data on the Internet has become increasingly important.Traditional sentiment analysis generally divides text into two different types based on sentiment expressed by the text,but this kind of coarse-grained sentiment analysis is increasingly unable to meet people's needs,such as "Apple pie is delicious,but the waiter The attitude is too bad."In response to this restaurant review,the emotional polarity is obviously different for the two attributes of Apple Pie "and"Service",but coarse-grained sentiment analysis can only give a rough sentiment judgment,it is unable to meet people's needs,and fine-grained sentiment analysis for specific entities can solve this problem well.This topic is based on deep learning methods.By studying the fine-grained sentiment analysis baseline model,we find the shortcomings and deficiencies in the model.With reference to its follow-up work,we combine our innovation to build a suitable deep learning model to achieve fine-grained sentiment analysis of product reviews..The specific research content of this topic is as follows.(1)For a given entity,to analyze the emotional attitude of product reviews,it is very important to consider how to interact information between the entity and the product reviews.This topic analyzes the previous work and finds its advantages and disadvantages.It adopts the method of establishing the connection between the two by establishing an aspect-context fusion layer.This method mainly uses the idea of holographic reduced representations to realize the entity and the commodity.The information in the comments blends.(2)This topic introduces interactive attention mechanism,models entities and comments separately and adopts attention mechanism to further realize the information fusion of entities and comments.In the attention mechanism here,the query vector isn't randomly initialized.The comment text vector is used as the query vector when the attention mechanism is used for the entity,the entity vector is used as the query vector when the attention mechanism is used for the review text.contact.(3)Use the Transformer model proposed in "Attention Is All You Need" to replace the LSTM model as a feature extractor.For fine-grained sentiment analysis,the self-attention mechanism is mainly to generate a vector representation of each word.When encoding each word,focus on the other parts of the input sentence to generate a word vector for each word.The vector here represents the output of the hidden layer similar to the LSTM model,including sentence semantics.Compared with the previous models such as CNN and RNN,the transformer model has the following advantages:low computational complexity;parallel computing;solving the long sequence dependency problem;it is a good model choice for fine-grained sentiment analysis.In this project,the LSTM model is replaced by a transformer model,and an effective fusion of the entity and the review text is combined with the attention mechanism to form the final model framework.